Tailing coal ash content detection method and system based on online class incremental learning

Abstract

The invention discloses a tail coal ash content detection method and system based on online class incremental learning, and the method comprises the steps: obtaining a coal slime mixed liquid data set, and carrying out the data preprocessing, and obtaining the two-dimensional spectrum image data of the coal slime mixed liquid; an online class incremental learning mechanism is introduced, and a tail coal ash content detection network model based on class incremental learning is constructed based on a backbone neural network model; and introducing a serding strategy, and carrying out tail coal ash content detection on the two-dimensional spectral image data of the coal slime mixed liquid based on a class incremental learning tail coal ash content detection network model to obtain a tail coal ash content detection result of the coal slime mixed liquid. According to the method, disastrous forgetting of the model in the training process can be resisted, the time of the model is shortened, and then the tailing ash content detection precision of the coal slime mixed liquid is improved. The tailing ash content detection method and system based on online class incremental learning can be widely applied to the technical field of coal slime flotation ash content detection.

Description

Technical Field

[0001] The present invention relates to the technical field of coal slime flotation ash detection, and in particular to a tail coal ash detection method and system based on online incremental learning. Background Art

[0002] Coal slime flotation is an important means of coal production. The tail coal ash information not only reflects the ash data of clean coal, but also reflects the recovery rate information of clean coal. At present, the technical methods for tail coal ash detection are all aimed at getting rid of manual coal selection and improving the detection accuracy. For example, they rely on machine vision or spectral data to model specific deep learning models to achieve the detection function. However, the current tail coal ash detection methods do not take into account the concept drift caused by changes in various factors such as coal slime properties in the industrial field. Secondly, the existing methods can only detect coal samples with limited ash categories. When there are new categories of coal samples, the original model can only be discarded and then a new model can be constructed by re-collecting new data and the original data set. That is, in the current research methods for tail coal ash detection in coal slime flotation, the ash labels of new data and the ash labels obtained by training the model using historical data are invalid and mismatched due to factors such as coal slime properties, particle size composition, and mineral impurities. At the same time, the arrival of new data also needs to be integrated with valid old data and retrain the model, which consumes more computing resources. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a tail coal ash detection method and system based on online incremental learning, which can resist catastrophic forgetting of the model during training and reduce the time of the model, thereby improving the tail coal ash detection accuracy of coal slime mixture.

[0004] The first technical solution adopted by the present invention is: a method for detecting tail coal ash based on online incremental learning, comprising the following steps:

[0005] Acquire a coal slime mixture data set and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture;

[0006] Introducing online incremental learning mechanism, based on the backbone neural network model, constructing the network model for tail coal ash detection based on incremental learning;

[0007] The herding strategy is introduced, and the tail coal ash detection network model is used to detect the tail coal ash in the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning, and the tail coal ash detection result of the coal slime mixture is obtained.

[0008] Furthermore, the step of acquiring a coal slime mixture data set and performing data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture specifically includes:

[0009] Obtaining several groups of coal slime mixtures with different concentrations and different ash contents;

[0010] The coal slime mixture is driven by a peristaltic pump to form a liquid flow and a flow-through cuvette observation window is set to collect data, thereby obtaining an image data set of the coal slime mixture;

[0011] The absorption spectrum data of the coal slime mixture image data set is collected by combining the fiber optic spectrometer with the xenon lamp light source to obtain the spectrum time series data of the coal slime mixture image;

[0012] The spectral time series data of the coal slime mixture image is converted into two-dimensional image data by using the Gram angle field method to obtain the two-dimensional spectral image data of the coal slime mixture.

[0013] Further, the step of introducing the herding strategy, performing tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning tail coal ash detection network model, and obtaining the tail coal ash detection result of the coal slime mixture specifically includes:

[0014] The two-dimensional spectral image data of the coal slime mixture is divided and processed to obtain several classification data sets;

[0015] Inputting the first classification data set into the backbone neural network model based on the class incremental learning tail coal ash detection network model for training, and obtaining a trained backbone neural network model;

[0016] Obtain weight data and bias data based on the trained backbone neural network model;

[0017] The incremental learning tail coal ash content detection network model is updated according to the weight data and the bias data to obtain an updated incremental learning tail coal ash content detection network model;

[0018] Samples of the first classification data set are selected through the herding strategy to obtain a representative data set;

[0019] The second classification data set is combined with the representative data set and input into the updated incremental learning tail coal ash detection network model for training until all classification data sets are traversed to obtain the final incremental learning tail coal ash detection network model;

[0020] Based on the final incremental learning tail coal ash detection network model, the tail coal ash is detected on the two-dimensional spectral image data of the coal slime mixture to obtain the tail coal ash detection result of the coal slime mixture.

[0021] Furthermore, the step of inputting the first classification data set into the backbone neural network model based on the class incremental learning tail coal ash detection network model for training to obtain the trained backbone neural network model specifically includes:

[0022] Inputting the first classification data set into a backbone neural network model based on a class incremental learning tail coal ash detection network model, wherein the backbone neural network model comprises an input layer, an initial convolutional layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer and a fully connected layer;

[0023] Based on the input layer of the backbone neural network model, obtaining a first classification data set;

[0024] Based on the initial convolutional layer of the backbone neural network model, preliminary feature extraction processing is performed on the first classification data set to obtain low-level features of the first classification data set;

[0025] Based on the first residual layer, the second residual layer, the third residual layer and the fourth residual layer of the backbone neural network model, deep feature extraction processing is performed on the low-level features of the first classification data set to obtain the deep features of the first classification data set;

[0026] Based on the fully connected layer of the backbone neural network model, the deep features of the first classification data set are linearly transformed and mapped to obtain the detection results of the first classification data set;

[0027] Determining the detection accuracy of the detection result of the first classification data set;

[0028] If the detection accuracy of the detection result of the first classification data set does not meet the preset accuracy threshold, the first classification data set is re-input into the backbone neural network model for training until the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold;

[0029] If the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold, the trained backbone neural network model is output.

[0030] Further, the step of updating the incremental learning tail coal ash detection network model according to the weight data and the bias data to obtain the updated incremental learning tail coal ash detection network model specifically includes:

[0031] For the backbone neural network model, generate a new fully connected layer;

[0032] Copy the number of output categories of the fully connected layer of the original backbone neural network model to the new fully connected layer;

[0033] Assign values ​​to the new fully connected layer according to the weight data and the bias data to obtain the assigned fully connected layer;

[0034] The assigned fully connected layer is replaced with the original fully connected layer to obtain an updated incremental learning network model for tail coal ash detection.

[0035] Furthermore, the expression for updating the incremental learning tail coal ash detection network model according to the weight data and the bias data is specifically as follows:

[0036] fc=update(fc,generate_fc(F,nb_classes))

[0037] In the above formula, represents the new fully connected layer, Represents the feature dimension.

[0038] Furthermore, the step of selecting samples from the first classification data set by using the herding strategy to obtain a representative data set specifically includes:

[0039] Get the current set embedding category center of the first classification data set;

[0040] Calculate the distance between each instance in the first classification data set and the center of the current set embedding category, and sort them in ascending order to obtain the sorted distance result;

[0041] The distance results after arrangement are selected according to the preset number of representative samples to obtain a representative data set.

[0042] Furthermore, the expression for obtaining the current set embedding category center of the first classification data set is specifically as follows:

[0043]

[0044] In the above formula, y represents the category of the first classification data set, n represents the total number of instances of the first classification data set, and X i represents the first classification data set, i represents the i-th instance, μ y Represents the current collection embedding category center.

[0045] The second technical solution adopted by the present invention is: a tail coal ash detection system based on online incremental learning, comprising:

[0046] The first module is used to obtain a coal slime mixture data set and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture;

[0047] The second module is used to introduce the online incremental learning mechanism and build a network model for tail coal ash content detection based on incremental learning based on the backbone neural network model;

[0048] The third module is used to introduce the herding strategy and perform tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning tail coal ash detection network model to obtain the tail coal ash detection result of the coal slime mixture.

[0049] The beneficial effects of the method and system of the present invention are as follows: the present invention obtains a coal slime mixture data set and performs data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture, and then introduces an online incremental learning mechanism, and based on a backbone neural network model, constructs a network model for tail coal ash detection based on incremental learning, and data is continuously input into the incremental model in the form of a data stream, and the model is continuously updated. Through the online incremental learning mechanism, while preserving the weight data and classification labels of the first coal sample training, incremental training can be performed on the data set of the second coal sample, so as to avoid retraining the model and shorten the training time of the model, and then introduces a herding strategy, and performs tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the incremental learning tail coal ash detection network model, and adopts the herding strategy to select a small number of representative samples from the previous task data subset and put them into the current training data set, which can resist catastrophic forgetting of the model during the training process while ensuring that the memory cost is not too high, thereby improving the detection accuracy of tail coal ash in the coal slime mixture. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the steps of a method for detecting ash content of tail coal based on online incremental learning of the present invention;

[0051] Figure 2 It is a structural block diagram of a tail coal ash detection system based on online incremental learning of the present invention;

[0052] Figure 3 It is a schematic diagram of a tail coal ash detection model based on online incremental learning provided by a specific embodiment of the present invention;

[0053] Figure 4 It is a structural schematic diagram of a backbone neural network model provided by a specific embodiment of the present invention;

[0054] Figure 5 It is a schematic diagram of online class incremental learning training provided by a specific embodiment of the present invention.

[0055] Figure numerals: 1, input layer; 2, initial convolution layer; 3, first residual layer; 4, second residual layer; 5, third residual layer; 6, fourth residual layer; 7, fully connected layer. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0057] First, it is necessary to explain the technical terms in the embodiments of the present invention:

[0058] 1) Coal slime flotation: Coal slime is a byproduct produced during the operation of coal washing plants. It is a semi-solid mixture of coal powder and water. However, the quality of coal is different, and the quality and availability of coal slime vary greatly. There are many types of coal slime and its uses are wide.

[0059] 2) Class-Incremental Learning: Class-Incremental Learning is a machine learning paradigm, also called continuous learning, which allows the model to retain memories of previously learned categories while continuously learning new categories. This learning method simulates the ability of humans to retain old memories when learning new things. In class-incremental learning, the model can only be exposed to the categories of the current task in each task, but not the categories of the previous tasks. This requires the model to effectively learn new categories without forgetting the old categories. Class-incremental learning is defined as Class-incremental learning aims to learn new features from a continuously updated data stream and resist catastrophic forgetting of old features. Suppose there are B training task sequences with no overlapping categories {D 1 ,D 2 ,D 3 ,…,D B},in refers to the data set of the bth incremental task, is category y i ∈Y b An example data, Y b is the label space of task b, where We can only access data set D when training task b. b The ultimate goal of incremental learning is to build a complete classification recognition model for all data categories. That is, the model cannot accurately recognize the current task set D b The corresponding categories can also accurately identify the categories corresponding to previous tasks. After each stage of the task, the model will evaluate its performance based on all the visible categories it has trained on.

[0060] 3) Concept drift: Concept drift is an important concept in the field of machine learning and data mining. It describes the phenomenon that the relationship between input data and target variables (labels) changes over time. This change means that the model originally learned from historical data may no longer be effective in predicting new data in the future because the relationship between input and output data has changed.

[0061] Reference Figure 1The present invention provides a method for detecting tail coal ash based on online class incremental learning, the method comprising the following steps:

[0062] S100, acquiring a coal slime mixture data set and performing data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture;

[0063] Specifically, several groups of coal slime mixtures with different concentrations and different ash contents are obtained; the coal slime mixture is driven by a peristaltic pump to form a liquid flow and a flow cuvette observation window is set to collect data, so as to obtain a coal slime mixture image data set; an optical fiber spectrometer and a xenon lamp light source are used to collect absorption spectrum data of the coal slime mixture image data set to obtain spectral time series data of the coal slime mixture image; the spectral time series data of the coal slime mixture image is converted into two-dimensional image data by the Gram angle field method to obtain two-dimensional spectral image data of the coal slime mixture.

[0064] In this embodiment, first, a total of 110 groups of coal slime mixtures with different concentrations and different ash contents are configured to simulate the flotation site, with a total of 110 groups, 11 types of ash, and 10 concentrations of each type of ash. A peristaltic pump is further used to drive the mixed liquid to form a liquid flow, and a flow cuvette observation window is set on the liquid flow branch for data collection. A micro UV-Vis fiber optic spectrometer is further used with a xenon lamp light source to continuously collect absorption spectrum data of the suspension liquid flow. Finally, the collected absorption spectrum time series data is preprocessed, and the Gram angle field technology is applied to convert the time series data into two-dimensional image data, and the converted data is shuffled and divided into multiple task subsets for subsequent training.

[0065] In addition, it should be noted that when used on industrial sites, a branch line can be set up next to the flotation machine, and a fiber optic spectrometer can be used to collect real-time spectral data of the liquid flow on the branch line.

[0066] S200, introduce online incremental learning mechanism, and build a network model for tail coal ash content detection based on incremental learning based on the backbone neural network model;

[0067] S300, introduce the herding strategy, perform tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning tail coal ash detection network model, and obtain the tail coal ash detection result of the coal slime mixture.

[0068] First of all, it should be noted that Figure 3As shown, the total data set is randomly divided into multiple task data subsets, where sub-data set 1 = data subset 1. After the model is initialized (in the PyTorch architecture, the corresponding weights and biases of the model will be automatically initialized), sub-data set 1 will be input into the model for initial training. After the first training, if the model performance does not meet the expected standard, it is necessary to stop the training and readjust. If the model performance meets the expected standard or above, the corresponding model parameter weights will be stored for subsequent use and enter the incremental training stage.

[0069] S310, dividing and processing the two-dimensional spectral image data of the coal slime mixture to obtain a plurality of classification data sets;

[0070] S320, inputting the first classification data set into the backbone neural network model based on the class incremental learning tail coal ash detection network model for training, to obtain a trained backbone neural network model;

[0071] Specifically, the first classification data set is input into the backbone neural network model based on the class incremental learning tail coal ash detection network model, and the backbone neural network model includes an input layer 1, an initial convolution layer 2, a first residual layer 3, a second residual layer 4, a third residual layer 5, a fourth residual layer 6 and a fully connected layer 7; based on the input layer of the backbone neural network model, the first classification data set is obtained; based on the initial convolution layer of the backbone neural network model, the first classification data set is subjected to preliminary feature extraction processing to obtain the low-level features of the first classification data set; based on the first residual layer, the second residual layer, the third residual layer and the fourth residual layer of the backbone neural network model, the low-level features of the first classification data set are subjected to deep feature extraction processing to obtain the low-level features of the first classification data set. The invention relates to a method for extracting and processing layer features to obtain deep features of the first classification data set; based on the fully connected layer of the backbone neural network model, linear transformation and mapping are performed on the deep features of the first classification data set to obtain the detection result of the first classification data set; the detection accuracy of the detection result of the first classification data set is judged; if the detection accuracy of the detection result of the first classification data set does not meet the preset accuracy threshold, the first classification data set is re-input into the backbone neural network model for training until the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold; if the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold, the trained backbone neural network model is output.

[0072] In this embodiment, if Figure 4 As shown in the figure, the incremental learning framework uses the ResNet18 network model as an example for training. The specific training process is as follows:

[0073] Input layer: The input image size is processed into a standard size of 224×224 pixels.

[0074] Initial convolution processing layer: Use a 7×7 convolution kernel to extract a large range of features from the input image, reduce the feature map resolution from 224×224 to 56×56 through convolution with a step size of 2 and 3×3 maximum pooling, and combine batch normalization and ReLU activation function to enhance the stability of the model and the initial feature expression ability, providing rich low-level features for subsequent layers.

[0075] The first residual layer: consists of two residual blocks, each of which contains two 3×3 convolution operations, with a convolution kernel size of 3×3, a padding of 1, a number of channels of 64, a step size of 1, and is followed by batch normalization and ReLU activation functions. The input is directly added to the output through the residual connection to ensure that the input feature information is retained while extracting shallow features, thereby alleviating the gradient vanishing problem and improving the optimization stability and training efficiency of the model.

[0076] The second residual layer: consists of two residual blocks, and the number of channels increases from 64 to 128. The first residual block halves the feature map size by setting the stride of the first layer of convolution to 2, and uses 1×1 convolution to adjust the number of channels in the residual connection so that the input dimension matches the output; the second residual block keeps the feature map size unchanged and continues to perform 3×3 convolution to extract features. Each residual block includes two 3×3 convolutions, followed by batch normalization and ReLU activation function, and the residual connection retains feature information. The main purpose of this layer is to extract deeper intermediate features by downsampling and increasing the number of channels, while retaining the input features to ensure training stability.

[0077] The third residual layer: It consists of two residual blocks, and the number of channels increases from 128 to 256. The first residual block halves the size of the feature map by setting the stride of the first layer of convolution to 2, and adjusts the number of channels through 1×1 convolution in the residual connection to match the input dimension with the output; the second residual block keeps the size and number of channels of the feature map unchanged, and continues to use 3×3 convolution to extract features. Each residual block contains two 3×3 convolutions, each followed by batch normalization and ReLU activation function. The residual connection retains the input feature information within each block to avoid gradient disappearance. The main purpose of this layer is to extract deeper mid- and high-level features, further enhance the network's ability to express complex patterns, and improve computational efficiency by reducing the resolution of the feature map.

[0078] The fourth residual layer: It consists of two residual blocks, and the number of channels is increased from 256 to 512, which is used to extract high-level semantic features. The first layer of 3×3 convolution of the first residual block halves the feature map size by setting the stride to 2, and adjusts the number of input channels in the residual connection to match the number of output channels through 1×1 convolution; the second residual block keeps the feature map size unchanged, and continues to extract deep features through two 3×3 convolutions. The convolution operation of each residual block is followed by batch normalization and ReLU activation function, and the input features are directly added through the residual connection to avoid the gradient vanishing problem and retain important information. This layer enables the network to capture complex patterns by increasing the number of channels and compressing the resolution of the feature map, providing strong feature support for the final classification task.

[0079] Fully connected layer: The fully connected layer maps the 512-dimensional feature vector generated by global average pooling to the number of categories of the classification task through linear transformation, and generates category probability distribution through the Softmax activation function to complete the final classification output, while reducing the number of parameters to reduce the risk of overfitting.

[0080] Finally, it should be noted that the use of the classic neural network ResNet18 as the backbone network has the advantage of maintaining high detection accuracy while having relatively few network parameters, making it easier to deploy in industrial field hardware platforms. In addition, ResNet18 has a large number of pre-trained models available, which makes it possible to quickly start the training process through transfer learning in incremental learning tasks, accelerate convergence and improve initial performance.

[0081] When the model is evaluated (i.e. the number of training rounds is reached and the model performance meets the standard), the next incremental learning training cycle can be carried out until all category tasks in the total data set are updated, that is, the training is completed.

[0082] S330, obtaining weight data and bias data based on the trained backbone neural network model;

[0083] S340, updating the incremental learning tail coal ash detection network model according to the weight data and the bias data to obtain an updated incremental learning tail coal ash detection network model;

[0084] Specifically, for the backbone neural network model, a new fully connected layer is generated; the number of output categories of the fully connected layer of the original backbone neural network model is copied to the new fully connected layer; the new fully connected layer is assigned according to the weight data and the bias data to obtain the assigned fully connected layer; the assigned fully connected layer is replaced with the original fully connected layer to obtain an updated class incremental learning tail coal ash detection network model.

[0085] In this embodiment, the first step of the incremental training phase is to update the model, that is, update the fully connected layer: retain the weights and biases learned in the previous training, while allowing the model to adapt to the new number of categories. If the number of categories increases, the new weights and biases will be initialized (usually random values), while the old weights and biases remain in their original positions.

[0086] First, a new fully connected layer is generated, and its expression is:

[0087] fc=generate_fc(F,C)

[0088] Where F is the feature dimension and C is the number of new categories.

[0089] Assuming nb_output is the number of output categories of the original fully connected layer, copy the original weights and biases to the first nb_output elements of the new fully connected layer, and the expression is:

[0090] W new [0:nb_output] = W

[0091] b new [0:nb_output] = b

[0092] Then delete the old fully connected layer and assign the new fully connected layer to fc, which is expressed as:

[0093] fc=fc new

[0094] The entire update process can be expressed as:

[0095] fc=update(fc,generate_fc(F,nb_classes))

[0096] This process ensures that the weights and biases of the new fully connected layers are properly preserved and updated when adding classes.

[0097] After the model is updated and the herding strategy completes the selection of representative data sets, the sub-data sets will enter the updated model for further training.

[0098] S350, selecting samples from the first classification data set by using a herding strategy to obtain a representative data set;

[0099] Specifically, the current set embedding category center of the first classification data set is obtained; the distance from each instance in the first classification data set to the current set embedding category center is calculated, and the distance results after arrangement are obtained by arranging them in ascending order; the distance results after arrangement are selected according to a preset number of representative samples to obtain a representative data set.

[0100] In this embodiment, in each incremental task, the model can generally only access the current corresponding data set, which is conducive to protecting user privacy and reducing storage burden. However, in order to resist catastrophic forgetting while ensuring that the memory cost is not too high, the embodiment of the present invention adopts a herding strategy to select a small number of representative samples from the previous task data subset and put them into the current training data set: Given a data set X from category y = {x 1 ,x 2 ,x 3 ,…,x n}, the herding strategy first calculates the category center of the current set embedding:

[0101]

[0102] Then calculate the distance from each instance to the category center And arrange them in ascending order, and select the most appropriate sample based on the ranking.

[0103] In addition, it should be noted that the embodiment of the present invention limits the number of representative samples: the number of representative samples of each category = 2000 / the current number of categories.

[0104] Therefore, the data set in the current incremental training stage is composed of: sub-dataset i = data subset i + representative data set, where data subset i is the sub-dataset divided by the total data set, and the representative data set is a collection of representative samples selected by the herding strategy.

[0105] S360, combining the second classification data set with the representative data set and inputting it into the updated class incremental learning tail coal ash detection network model for training, until all classification data sets are traversed to obtain the final class incremental learning tail coal ash detection network model;

[0106] S370. Based on the final class-incremental learning tail coal ash detection network model, tail coal ash detection is performed on the two-dimensional spectral image data of the coal slime mixture to obtain the tail coal ash detection result of the coal slime mixture.

[0107] Finally, the specific implementation steps of the present invention are summarized. The embodiment of the present invention is a method that can dynamically process coal sample data online and gradually update. First, data is continuously input into the class incremental model in the form of data stream, and the model is continuously updated. Historical data corresponds to historical labels, and new data corresponds to new labels. While learning new data, old data is not so easy to be forgotten, that is, the ash content of previous coal samples and the ash content of newly added coal samples can be effectively detected. The model is continuously updated according to the current coal sample quality to meet industrial needs.

[0108] like Figure 5As shown, the embodiment of the present invention applies the class incremental learning technology to the ash content detection of coal slime flotation tailings, and can be divided into different task stages according to the ash categories that actually need to be detected. For example, a coal sample originally having ten categories of ash can be put into Task 1 to train a ten-category model, and all current ash categories can be detected. When there are coal samples with new ash categories in the new stage, they can be used as Task 2 to continue training the original ten-category model and update the categories at the same time, without the need to retrain a new model with all the previous coal samples and new coal samples.

[0109] Finally, it should be explained that the number of times the model needs to be trained for the entire class incremental learning framework of the embodiment of the present invention depends on the number of tasks (coal sample changes). Specifically, assuming that this round of coal slime flotation production contains three batches of different coal samples, first collect the absorption spectrum data of the first coal sample, train the ResNet18 classification model, and identify the 10 types of ash in this batch of coal samples. However, when it comes to the second batch of coal samples, their configuration and other element contents may be different from those of the first batch of coal samples, which leads to the extremely poor classification effect of the first trained model on the second batch of samples. However, it is time-consuming and labor-intensive to start training from the beginning for the second batch of coal samples, and the trained model can only be used for a certain batch of coal samples. Therefore, the framework effect of class incremental learning is reflected. While saving the weight data and classification labels of the first coal sample training, incremental training is performed on the data set of the second coal sample. While maintaining the original 10 classifications, 10 new classification labels suitable for the second batch of coal samples are added. When the third batch of coal samples flows into production, configuration changes may occur again. Incremental training is performed according to the above description to avoid starting training from the beginning. After several days of production, if the probability of a certain ash content in a certain coal sample appearing is extremely small during model training or recognition, the model will selectively forget this label to avoid model redundancy.

[0110] Reference Figure 2 , a tail coal ash detection system based on online incremental learning, comprising:

[0111] The first module 201 is used to obtain a coal slime mixture data set and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture;

[0112] The second module 202 is used to introduce an online incremental learning mechanism and build a tail coal ash detection network model based on incremental learning based on the backbone neural network model;

[0113] The third module 203 is used to introduce a herding strategy, and perform tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning tail coal ash detection network model to obtain the tail coal ash detection result of the coal slime mixture.

[0114] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0115] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for detecting tail coal ash content based on online incremental learning, characterized in that: The following steps are involved: Acquire a coal slime mixture data set and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture; Introducing online incremental learning mechanism, based on the backbone neural network model, constructing the network model for tail coal ash detection based on incremental learning; The herding strategy is introduced, and the tail coal ash detection network model is used to detect the tail coal ash in the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning, and the tail coal ash detection result of the coal slime mixture is obtained.

2. According to claim 1, a method for detecting tail coal ash based on online incremental learning is characterized in that: The step of acquiring a coal slime mixture data set and performing data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture specifically includes: Obtaining several groups of coal slime mixtures with different concentrations and different ash contents; The coal slime mixture is driven by a peristaltic pump to form a liquid flow and a flow-through cuvette observation window is set to collect data, thereby obtaining an image data set of the coal slime mixture; The absorption spectrum data of the coal slime mixture image data set is collected by combining the fiber optic spectrometer with the xenon lamp light source to obtain the spectrum time series data of the coal slime mixture image; The spectral time series data of the coal slime mixture image is converted into two-dimensional image data by using the Gram angle field method to obtain the two-dimensional spectral image data of the coal slime mixture.

3. According to claim 2, a method for detecting tail coal ash based on online incremental learning, characterized in that: The step of introducing the herding strategy, performing tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning tail coal ash detection network model, and obtaining the tail coal ash detection result of the coal slime mixture specifically includes: The two-dimensional spectral image data of the coal slime mixture is divided and processed to obtain several classification data sets; Inputting the first classification data set into the backbone neural network model based on the class incremental learning tail coal ash detection network model for training, and obtaining a trained backbone neural network model; Obtain weight data and bias data based on the trained backbone neural network model; The incremental learning tail coal ash content detection network model is updated according to the weight data and the bias data to obtain an updated incremental learning tail coal ash content detection network model; Samples of the first classification data set are selected through the herding strategy to obtain a representative data set; The second classification data set is combined with the representative data set and input into the updated incremental learning tail coal ash detection network model for training until all classification data sets are traversed to obtain the final incremental learning tail coal ash detection network model; Based on the final incremental learning tail coal ash detection network model, the tail coal ash detection is performed on the two-dimensional spectral image data of the coal slime mixture to obtain the tail coal ash detection result of the coal slime mixture.

4. According to claim 3, a method for detecting tail coal ash based on online incremental learning is characterized in that: The step of inputting the first classification data set into the backbone neural network model based on the class incremental learning tail coal ash detection network model for training to obtain the trained backbone neural network model specifically includes: Inputting the first classification data set into a backbone neural network model based on a class incremental learning tail coal ash detection network model, wherein the backbone neural network model comprises an input layer, an initial convolutional layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer and a fully connected layer; Based on the input layer of the backbone neural network model, a first classification data set is obtained; Based on the initial convolutional layer of the backbone neural network model, preliminary feature extraction processing is performed on the first classification data set to obtain low-level features of the first classification data set; Based on the first residual layer, the second residual layer, the third residual layer and the fourth residual layer of the backbone neural network model, deep feature extraction processing is performed on the low-level features of the first classification data set to obtain the deep features of the first classification data set; Based on the fully connected layer of the backbone neural network model, the deep features of the first classification data set are linearly transformed and mapped to obtain the detection results of the first classification data set; Determining the detection accuracy of the detection result of the first classification data set; If the detection accuracy of the detection result of the first classification data set does not meet the preset accuracy threshold, the first classification data set is re-input into the backbone neural network model for training until the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold; If the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold, the trained backbone neural network model is output.

5. According to claim 4, a method for detecting tail coal ash based on online incremental learning is characterized in that: The step of updating the incremental learning tail coal ash detection network model according to the weight data and the bias data to obtain the updated incremental learning tail coal ash detection network model specifically includes: For the backbone neural network model, generate a new fully connected layer; Copy the number of output categories of the fully connected layer of the original backbone neural network model to the new fully connected layer; Assign values ​​to the new fully connected layer according to the weight data and the bias data to obtain the assigned fully connected layer; The assigned fully connected layer is replaced with the original fully connected layer to obtain an updated incremental learning network model for tail coal ash detection.

6. The method for detecting tail coal ash based on online incremental learning according to claim 5, characterized in that: The expression for updating the incremental learning tail coal ash detection network model according to the weight data and the bias data is specifically as follows: ； In the above formula, represents the new fully connected layer, Represents the feature dimension.

7. The method for detecting tail coal ash based on online incremental learning according to claim 6, characterized in that: The step of selecting samples from the first classification data set by using the herding strategy to obtain a representative data set specifically includes: Get the current set embedding category center of the first classification data set; Calculate the distance between each instance in the first classification data set and the center of the current set embedding category, and sort them in ascending order to obtain the sorted distance result; The distance results after arrangement are selected according to the preset number of representative samples to obtain a representative data set.

8. The method for detecting tail coal ash based on online incremental learning according to claim 7, characterized in that: The expression for obtaining the current set embedding category center of the first classification data set is specifically as follows: ； In the above formula, Represents the category of the first classification data set, represents the total number of instances of the first classification dataset, represents the first classification data set, Indicates Instances, Represents the current collection embedding category center.

9. A tail coal ash detection system based on online incremental learning, characterized in that: Includes the following modules: The first module is used to obtain a coal slime mixture data set and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture; The second module is used to introduce the online incremental learning mechanism and build a network model for tail coal ash content detection based on incremental learning based on the backbone neural network model; The third module is used to introduce the herding strategy and perform tail coal ash detection on the two-dimensional spectral image data of the coal slime mixture based on the class incremental learning tail coal ash detection network model to obtain the tail coal ash detection result of the coal slime mixture.

Images